world models · rl environments
Your agents should fail in a dream, not in production.
Oneirology builds reinforcement-learning environments out of the systems you already run. Agents rehearse thousands of rollouts inside a model of your world — and every consequential action is checked against the real thing, so what an agent learns is what's actually true.
Record what actually changes
We capture the state your systems change: the rows written, the resources created, the queues drained. That record is the ground truth everything downstream is measured against.
Compile the graders
Someone demonstrates a workflow once. The task and its pass/fail checks compile out of the state difference — no hand-written assertions, and no model deciding whether the work got done.
Learn a model of the world
We fit a transition model to those traces: the states your systems occupy, the actions available, and what actually happens next.
Rehearse, then verify
Agents run thousands of rollouts inside the model. The cheap steps are imagined; the ones that change something are checked against real state, so a policy can't learn to exploit a hallucination.
An LLM agent's score on a task no human wrote, graded from stored state rather than from what the agent claimed.
Of 6,775 assertions across 100 hand-written verifiers, the share that read state a before-and-after diff can see.
Explored actions that changed anything at all — which is why a demonstration, not a crawler, is the input.
Research
Reading Frozen World Models Through Action Effects
Interpretability work on action-conditioned world models usually examines the latent state. This asks the complementary question: how does a frozen predictor's output move under controlled actions? Joint approximate diagonalization of centered per-action effect covariances, followed by action-profile clustering, surfaces candidate action-selective subspaces without using state-factor labels — evaluated on sprite JEPA, V-JEPA-2-AC, and DINO-WM.
Read the paper →